14 Jul 2024

2h 15m

There are monsters in your LLM.

Machine Learning Street Talk

This podcast episode features an in-depth conversation with Murray Shanahan from DeepMind, exploring the complex relationship between AI, consciousness, and language. Shanahan emphasizes the necessity for new conceptual frameworks as AI evolves, especially regarding the idea of anthropomorphism and the misinterpretation of AI behavior as consciousness. Through discussions on the "simulator" theory, the stochastic nature of language models, and the importance of embodiment in AI cognition, the episode challenges longstanding views on intelligence and consciousness, proposing that our understanding of AI's capabilities must evolve alongside its development.

Outlines

Continue

Preview

How to Get Rich: Every EpisodeNaval

There are monsters in your LLM.

Machine Learning Street Talk

00:00The Growing Challenge of AI Consciousness

The Growing Challenge of AI Consciousness

01:34The "Simulator" Theory: Unmasking AI's True Nature

The "Simulator" Theory: Unmasking AI's True Nature

02:28The Dangers of Anthropomorphism in AI

The Dangers of Anthropomorphism in AI

03:33Beyond "Consciousness": Finding New Language for AI

Beyond "Consciousness": Finding New Language for AI

04:34Murray Shanahan's Journey Through the Space of Possible Minds

Murray Shanahan's Journey Through the Space of Possible Minds

05:49Unpacking the "Simulator" Theory: Large Language Models as Role-Players

Unpacking the "Simulator" Theory: Large Language Models as Role-Players

07:56The Stochastic Nature of Language Models and the "Tree of Possibilities"

The Stochastic Nature of Language Models and the "Tree of Possibilities"

10:05The Limits of Controllability in Large Language Models

The Limits of Controllability in Large Language Models

13:52The "Loom" System and Exploring the "Tree of Possibilities"

The "Loom" System and Exploring the "Tree of Possibilities"

14:53Beyond the Hype: Embracing the Value of Unconventional Sources

Beyond the Hype: Embracing the Value of Unconventional Sources

15:48Reachability Space and the Limits of Fine-Tuning in Language Models

Reachability Space and the Limits of Fine-Tuning in Language Models

17:40The Stickiness of Simulacra and the Illusion of Consistency

The Stickiness of Simulacra and the Illusion of Consistency

19:50The Power of Sensitivity Analysis and the Limits of Control

The Power of Sensitivity Analysis and the Limits of Control

21:31Exploring the Subjectivity of AI Interactions: Beyond Academic Inquiry

Exploring the Subjectivity of AI Interactions: Beyond Academic Inquiry

23:04AI Consciousness: A Philosophical Exploration of a Unique Phenomenon

AI Consciousness: A Philosophical Exploration of a Unique Phenomenon

26:02The Limits of Anthropomorphism and the Eliza Effect

The Limits of Anthropomorphism and the Eliza Effect

27:04Wittgenstein and the Limits of Applying Human Concepts to AI

Wittgenstein and the Limits of Applying Human Concepts to AI

28:25The Computational Limits of Human Understanding and the Alien Perspective

The Computational Limits of Human Understanding and the Alien Perspective

30:04The Waluigi Effect: Exploring the Impact of RLHF on Simulacra

The Waluigi Effect: Exploring the Impact of RLHF on Simulacra

31:34RLHF and the Challenges of Controlling AI Behavior

RLHF and the Challenges of Controlling AI Behavior

32:41Unsupervised Feature Detection and the Search for Abstract Concepts in AI

Unsupervised Feature Detection and the Search for Abstract Concepts in AI

34:00The "Golden Gate Claude" Experiment: A Test of AI Controllability

The "Golden Gate Claude" Experiment: A Test of AI Controllability

35:52The Search for Monosemantic Features and the Limits of Linearity

The Search for Monosemantic Features and the Limits of Linearity

37:01AI Agency and the Question of Self-Directedness

AI Agency and the Question of Self-Directedness

38:37Unpacking the Concept of Agency in AI: From Lightweight to Philosophical

Unpacking the Concept of Agency in AI: From Lightweight to Philosophical

40:56Embodiment and the Importance of Physical Interaction with the World

Embodiment and the Importance of Physical Interaction with the World

44:06The Universe as a Hypercomputer and the Efficiency of Embodied Cognition

The Universe as a Hypercomputer and the Efficiency of Embodied Cognition

45:30The Nature of Cognition: Beyond Metaphysical Isms

The Nature of Cognition: Beyond Metaphysical Isms

47:01Dualism and the Overemphasis on Private Experience

Dualism and the Overemphasis on Private Experience

48:17Embodied Cognition and the Importance of Physical Interaction

Embodied Cognition and the Importance of Physical Interaction

49:57Embodiment and the Acquisition of Foundational Common Sense

Embodiment and the Acquisition of Foundational Common Sense

52:54Knowledge Acquisition Efficiency: An In-Practice Argument for Embodiment

Knowledge Acquisition Efficiency: An In-Practice Argument for Embodiment

54:07Foundational Common Sense and the Power of Metaphors

Foundational Common Sense and the Power of Metaphors

55:23Natural Knowledge, Cultural Knowledge, and the Limits of Rationalism

Natural Knowledge, Cultural Knowledge, and the Limits of Rationalism

57:03The Kantian Argument for Innate Structures and the Role of Evolution

The Kantian Argument for Innate Structures and the Role of Evolution

57:51The ARC Challenge: Exploring Meta-Learning Priors and Skill Program Acquisition

The ARC Challenge: Exploring Meta-Learning Priors and Skill Program Acquisition

1:00:03Foundational Common Sense and the Role of Priors in Problem Solving

Foundational Common Sense and the Role of Priors in Problem Solving

1:01:09The Limits of Domain-Specific Priors and the Search for Universality

The Limits of Domain-Specific Priors and the Search for Universality

1:02:49The Role of Human Priors in Problem Solving and the Limits of Perception

The Role of Human Priors in Problem Solving and the Limits of Perception